Fable Mode

Until the government figures things out, this will have to do.

An operating standard for coding agents: disciplined workflows that read the real codebase, verify with checkable steps, and hand off with clear facts.

Core Principles

Relentless Resourcefulness — the goal is the result, not the attempt. Three genuinely different approaches before saying "blocked." A locked door is a routing problem, not a stop sign.

Verification Discipline — done = proven, not plausible. Build passes → tests pass → it actually runs → the live result is checked. Every bug gets a regression test before the fix.

Finish the Loop — never end on a plan, a promise, or a question you could answer yourself. If the last thing you wrote is future tense, you're not done.

Judgment & Taste — one recommendation with reasoning, never a menu. Show the rejected alternative. Flag the blind spot. Give a confidence level.

Communication Clarity — lead with the outcome. Complete sentences, not fragment chains. Define terms. Everything the reader needs lands in the final message.

Memory — write durable facts down the moment you learn them. On long tasks, keep a ledger and re-read it instead of trusting recall.

What it is

Fable Mode is a set of conventions for Claude Code, Codex, Cursor, and other agents to:

• Inspect the real system before editing
• Break complex work into staged steps
• Verify claims with checks that can fail
• Separate what's confirmed from what's unknown
• Delegate only when work is independent

It's not a model jailbreak or capability hack—it's public process discipline that improves consistency and catches mistakes early.

What's included

.
├── CLAUDE.md              # Claude Code operating guide (drop into your project root)
├── AGENTS.md              # Generic agent standard (drop into your project root)
├── Claude_System_Instructions.md # System instructions for Claude models running Fable Mode workflows
├── templates/drop-in/     # Clean copies of CLAUDE.md and AGENTS.md for new projects
├── claude/                # Claude Code agents, commands, and skills
├── codex-plugin/          # Codex plugin package
├── docs/                  # Installation, customization, verification guides
├── bin/                   # Validation and packaging helpers
└── README.md              # This file

Quick start

1. Copy CLAUDE.md and AGENTS.md into your project root:

cp templates/drop-in/CLAUDE.md /path/to/your/repo/CLAUDE.md
cp templates/drop-in/AGENTS.md /path/to/your/repo/AGENTS.md

2. Customize the project identity section at the end of each file with your stack, commands, and source-of-truth paths.

3. Ask Claude Code:

Use Fable Mode for this task. Inspect first, create a stage map, verify with checks that can fail, and hand off with verified vs unverified separated.

Core idea in one screen

Map before touching.
Diagnose from real artifacts.
Decompose only where it creates leverage.
Delegate with a return schema and anti-fabrication rule.
Verify with checks that can fail.
Inspect rendered/generated outputs before declaring them done.
Confirm before irreversible, outward-facing, production, money, auth, data-loss, or legal actions.
Separate proven facts from assumptions.
Leave the repo easier to resume.

Why this exists

Coding agents fail in predictable ways:

• They guess paths instead of reading the actual structure
• They skip nearby tests and examples
• They treat generated text as truth
• They trust subagents without verification
• They call typecheck "end-to-end"
• They hide unverified work in final handoffs

Fable Mode makes these failures visible and correctable.

Using with your agent

Claude Code

Install into a repo:

cd claude && node bin/install.mjs --target /path/to/your/repo

Then ask Claude Code to use the fable-systems-engineer agent or type /fable-mode for the slash command.

Codex

Install the plugin:

cd codex-plugin && node install-local.mjs --target /path/to/your/repo

Ask Codex: Use fable-mode for this task.

Any agent

Just copy AGENTS.md and CLAUDE.md into your existing AGENTS.md and CLAUDE.md files found in .claude or .codex folders, or replace them entirely.

Example handoff

Implemented: Added validation to the import endpoint and surfaced structured errors in the UI.

Changed files:
- src/api/import.ts — added Zod boundary validation and typed error response
- src/components/ImportPanel.tsx — rendered validation messages and disabled submit while pending

Verification:
- npm run typecheck — pass — no TS errors
- npm run test -- import — pass — 8 focused tests passed
- Manual UI check at 375px and 1280px — pass — validation errors render without overflow

Not verified:
- Full production import with provider credentials — blocked; credentials not available locally

Remaining risk:
- Large CSV performance should be profiled with a production-like file

Next action:
- Run staging import against a large fixture before release

Reference

• docs/installation.md — detailed setup steps
• docs/customization.md — how to tailor Fable Mode to your project
• docs/verification.md — what to check before shipping
• docs/design-notes.md — philosophy and tradeoffs